mcp2cli/mcp2skill

Also known as: mcp2skill · mcp-to-cli · mcp-to-skill · mcpcli Convert MCP servers into CLI and Agent SKILLs.

Why

For example, install zereight/gitlab-mcp — it exposes 122 tools across 14 categories (merge requests, issues, pipelines, wikis, etc.). Every tool schema gets injected into your AI conversation context on every request. We measured this directly against the Claude API:

After mcp2cli convert, your AI agent stops seeing 122 tool definitions. Instead, it reads one small skill file, and calls CLI commands like mcp2cli gitlab mr create --project-id 123 --title "Fix bug" to get the job done.

Beyond token savings, some projects maintain both a CLI and an MCP server in parallel (e.g. the GitLab MCP server vs. the glab CLI). Keeping two interfaces in sync is a maintenance burden and inevitably leads to feature drift. mcp2cli eliminates the need — one MCP server, one generated CLI, always consistent.

Quick Start

pip install mcp-to-cli

If you already have an MCP server configured in Claude, Cursor, or Codex:

mcp2cli list              # See which MCP servers are available
mcp2cli convert gitlab-mcp

That’s it. Your MCP server is now a CLI, and the compressed skill files are synced to your AI clients.

Tip: After install or convert, you can edit ~/.agents/mcp2cli/servers.yaml to update env vars (API keys, endpoints, etc.) without re-running the full setup.

How It Works

mcp2cli convert gitlab-mcp runs a 3-phase pipeline:

┌──────────────┐      ┌──────────────┐      ┌──────────────┐
│  1. Extract  │─────>│  2. Generate │─────>│   3. Sync    │
│              │      │              │      │              │
│  Read config │      │ AI-generate  │      │ Copy skills, │
│  from Claude,│      │ CLI + skill  │      │ disable MCP  │
│  Cursor,Codex│      │              │      │              │
└──────┬───────┘      └──────┬───────┘      └──────┬───────┘
       v                     v                     v
 mcp_servers.json    cli.yaml + SKILL.md    ~/.claude/skills/

1. Extract — Reads your MCP server config from Claude (~/.claude.json) / Cursor (~/.cursor/mcp.json) / Codex (~/.codex/config.toml)
2. Generate — Connects to the MCP server, discovers all tools, and uses AI to generate a CLI command tree (cli.yaml) and a compressed skill file (SKILL.md)
3. Sync — Copies skill files to your AI clients and disables the raw MCP server